More From The AI For Impact Class Feb. 15

I wrote about last week’s special guests in the previous posts, but here’s more on what happened in our class Thursday.

First, as I already mentioned, following our guest speaker’s suggestion to look at obscure verticals, Professor Ramesh Raskar had the class come up with a list:

· Immigration

· manufacturing procurement

· risk insurance modeling

· organizing railcars

· veterinary services

Then they talked about some of the challenges.

“(Sometimes) you don’t have the data visible,” Ramesh said. “So how are you going to crack that?”

He introduced the idea of “centralized” AI, giving the example of someone with a health condition trying to locate care geographically, and along with the class, enumerating four key challenges to solve: privacy, data accessibility, rewarding data contributors, and interpretability.

On the privacy front, Ramesh talked about possible models based on systems like Google Maps and Waze, where there’s a utility in giving up personal data in some form.

“Can we … get privacy and utility at the same time?” he asked. “If you can solve that one problem, we can probably solve almost every problem in the world: health, climate, transportation, agriculture, logistics, HR, possibly even democracy.”

Ramesh made the distinction between personal privacy, and something like national security. His anecdote about USSR spies looking at Pentagon pizza orders was interesting!

“Nobody cares about an individual’s privacy,” he said, speaking to a generalized case. “What they care about is something at the trade secret level, or the national security level, because somebody else is willing to buy the data from them.”

Ultimately, he concluded the first part of the lecture by stressing the significance of solving privacy-related challenges for unlocking the potential of AI in various sectors, with innovations contributing to global GDP.

And then there was part 2, where Ramesh went over anonymization, obfuscation, encryption (homomorphic encryption), differential privacy, federated learning, and split learning, also talking about privacy/utility trade-offs, the importance of scalability, and again, opportunities for entrepreneurship.

I thought it was interesting to look at some use cases he presented, which I’ll just go over briefly:

A decision support system for first responders – here we have some ideas for AR designs that can help police or others respond to emergency events

Immigrant assistance – using chatbots, designers could help immigrants who are filling out complicated forms useful in immigration law

Fostering financial inclusion in emerging markets – here we talked about financial access for small businesses in places like Africa, in supporting marginalized entrepreneurs, including women and those in less developed communities, with a focus on accommodating the unbanked, as well as the underbanked, both of which are quite large demographics

Workforce reskilling – we talked about the impact of automation on labor, and some startups that are helping make an impact

Government project management – ideas included a “copilot” system for project planning and problem solving

We also talked about active participation…

“Be the aggressive subset,” said Dave Blundin, encouraging students to grab the mic and share their ideas. “You’re going to get a lot of advice – and you’re going to learn to communicate incredibly effectively , by the end of the semester.”

Ramesh talked about zeroing in on good questions – which is very good advice.

At the end, I brought on former MIT Media Lab advisor Andy Karsner, who is now a senior strategist at Alphabet.

“I call this place the imagination factory” he said of MIT.

AI, he said, is “all we’re working on these days.” As for his bonafides, I think “FOJ” (Friend of John) is just a very minor part of a long list! (He has quite impressive government and private sector experience!)

Ok, that’s a lot of what we went over. Tune in next week!